Hacker News Reader

Comment by partiallypro | Hacker News Reader

I think there is a genuine concern that Google could overreact and launch a half-baked product in pure panic of being left behind or one upped by Microsoft. There is also a fear, I think, that the ChatGPT integration with Bing/Edge could not go all that smoothly. I think it could be game changing in many ways, but I can also see both of these falling apart.

Amazon Alexa, Google Assistant and Siri were thought to be good at launch, the press loved them...but now they are not nearly as valuable as they were touted (and actually lose these companies money.) Convenient, but not game changing. I think it's a waiting game to see what this truly does.

I do think there is unique break here though, because I feel that SEO has so thoroughly ruined search in many regards that this -could- be the right moment for this.

View on HN

> Amazon Alexa, Google Assistant and Siri were thought to be good at launch

In my own experience, Alexa and Google's query pucks were improving for a short time, and then got considerably worse, losing features every month until they basically stopped understanding or responding to anything but the simplest requests.

A couple months after released, they expanded android auto with the same abilities, so I could "ok google" in my car and ask for the answer to just about any question, or to adjust the temp in my house, or ask it to play just about any obscure album / artist. It improved over about 6 months and then began to devolve from there. It's not even possible to ask for "[popular artist] radio" anymore in the car nor can I run voice queries at all that aren't map-specific.

Not sure what happened there, but in my mind they failed miserably. I still like Android auto, but my google and alexa pucks are all in a pile in some cabinet around here somewhere.

I’m not sure Siri ever got as good as the original Siri app that Apple bought. Maybe the custom triggers are an advancement but it has spent most of its life chasing the original Siri.

Siri is such a disappointment. It fails on things that seem like they should be trivial like unit conversions. Pretty much only good for setting timers and making phone calls now.

> In my own experience, Alexa and Google's query pucks were improving for a short time, and then got considerably worse, losing features every month until they basically stopped understanding or responding to anything but the simplest requests.

Yes!! I thought it was just me and my Australian accent.

And the commands are super limited. I wonder to what extent that's also because third-party companies aren't integrating properly? Eg, my Roborock has clearly labeled rooms, but I can't say "hey Google, start vacuuming the kitchen". Is that Google's fault, or Roborock's?

Also, I can't seem to chain commands, eg "hey Google, set the lights to red and 10% brightness". I have to say them separately. That seems like a Google thing to me.

There was a period of 2 months when googles third party integration wouldnt let you add new items to an existing registered account.

So dumb.

FWIWm, Alexa and Roomba seem to work fine together, so I guess it depends on the quality of Roborock's intergration.

Years ago, before it was Google Assistant, my phone could hear me say OK Google while sitting in a cupholder driving down the road. Now my phone frequently fails to respond when I'm holding it near my face in a silent room. But somehow it still has more false positive activations. And they haven't added a single new feature I actually find useful in that entire time.

Sometimes retraining the voice model improves the response.

Personally, I also encounter clogged mic ports, which kills my OK google distance

I've tried retaining several times. It's certainly possible that my newer phones have poorly designed microphones. But all of the phones in question were Nexus/Pixel so it would still be Google's fault.

I've been so disapointed with Siri. First, the latency is just plain aweful. Second, I tried a number of times to ask it not to interrupt me with notifications but alas, it failed to understand. I had to find the setting and toggle it myself.

Apple has a lot of work to improve that. A lot of work...

I wonder if they nerfed their assistants because paths that are too autonomous too easily wander into PR nightmares. Carefully curated results and behavior isn't going to spontaneously generate content actual human beings would consider offensive or prejudiced. People are pretty eager to pull out their pitchforks, and "Big Tech Builds Vile Hate Bot" make a excellent click-bate headline every time.

The production of prejudiced content by "AI"s is a real issue and not just a problem of click-bait headlines

There is also the crap around regions. I'm in the UK, but if I set my assistant to the the US suddenly it can understand a lot more and perform a lot more tasks.

Although while it understands a lot more, it's still getting worse.

Same about my Fitbit Versa. It was great, and now Google removes features.

At this point I would just prefer it if they let me just replace the firmware with one of the opensource projects. Can I replace my Alexa or Google puck firmware with something better?

This is why I've kept my pucks. I'm hoping for a hackaday post or something where someone installs their own _something_ on them. They have great mics and the speakers are decent. Wouldn't mind using them as an interface to Mycroft or something.

Unfortunately Mycroft is dead.

Maybe Rhasspy with Home Assistant will save us from this mess.

>I do think there is unique break here though, because I feel that SEO has so thoroughly ruined search in many regards that this -could- be the right moment for this.

the problem here is monetization, will search even be profitable if LLMs are used for most queries? It might become a Uber/Lyft or food delivery situation where these companies aren't really able to profitably deliver the service. I don't see many people paying a subscription for search and there's no way governments will allow "native" advertising within answers without them being signaled as ads, which would hurt trust in responses

Microsoft might not care and just see it as a way to hurt Google's money printing machine and operate Bing at a loss or break even. Google Cloud and workspace are finished without Google's ad money funding them and Microsoft Azure and Office would gain

> the problem here is monetization, will search even be profitable if LLMs are used for most queries?

I think that is a concern, which is why I think Microsoft has a chance to just bundle this or a more advanced version with Microsoft 365. Then you're already paying for it. That automatically puts it in the hands of millions of paying customers and corporations. You can just raise your price by a dollar a month or something to offset the costs and no one really will even think they are paying for this.

That only makes sense if they can charge more for office 365.

Shoving more features into something people pay for doesn’t mean those feature were worth it.

Except that's not the calculus. By integrating chatgpt with their products, Microsoft pulls users away from: search, docs/workspace, and even gmail, and brings them over into: Bing, Office365, and outlook. In doing so, they threaten google's core income stream, and therefore google's ability to fight back.

Microsoft is not paying for an expensive search engine: Microsoft is paying (with chatgpt compute infrastructure) for a much larger piece of the productivity suite market, and hamstringing google in the process.

"Clippy, write me a 25 page requirements doc for an integration between Workday and our IAM solution."

Thanks to that joke I just imagined how it could be a game changer for internal corporate documents (the kind that are 50+ pages and never read), especially if it’s native into Office.

We'll have to invent a new form of busy work to justify head count.

Actually reading printed copies of those documents would be the penultimate busy works, so much that it would quickly become uberablingly alienating!

And what would be the ultimate busy works?

Employee: Chatbot, write me a 20 page document describing what I've been working on the past two weeks so that I can give it to my manager.

Manager: Chatbot, write me 1 page summaries of these documents from my employees so that I can collate them for the CEO.

CEO: Chatbot, where should I go on vacation next?

In Kino's Travels (I guess nowadays Kino's Journey), there was one country in the first anime adaptation where people no longer needed to work because the system/computer/whatever it was called did it all for them, so they spent all their time confirming its correctness.

I fucking LOVE Kino's Journey, nobody else I've met has ever seen it!

TIME TO REWATCH YET AGAIN

The advertising will probably be more insidious, but no less profitable. My guess is that the hidden pre-prompt will end up including something like:

>You are a generative model designed to provide reasonably correct information, with a preference for providing flattering portrayals of your advertising partners. Your advertising partners are ranked according to a token system...

This is easy to cross-check against the competing search engine automatically. Some third party may provide such a service.

Why would LLMs be less monetizable than search? People are still going to want to buy goods and services, and being the place they go to find out about those goods and services will be just as valuable as it is today. Perhaps even more so, since LLMs are able to answer questions that are difficult to formulate as search queries.

Because most queries are not shopping queries. Yet ads are injected regardless to tempt you into buying something.

Probably because they are too expensive. I heard it costs 7 cents per query on average for ChatGPT. That’s more than ads pay.

With the search query as the prompt, they could probably cache the response and use it several times. That might keep the cost down to something reasonable.

I'd go further and say that SEO has not only ruined search, but the internet itself.

There's so much crappy content around, generated webpages that aim to match every search query. I just looked for "four times five" and there's a huge ton of matches, including pages dedicated to "4x5" [0], which I think do more damage to the web than help.

I feel the web is broken. It was modeled around documents and slowly features started creeping in, like adding images, gifs, audio, video, applications (flash, java), 3D, making the document closer to a program. At every step it felt like an improvement, but it seems that we ended up with another C++. There's probably a law around creating a platform so popular that it'll organically get features until the accidental design is realized to be awful so people start coming up with their ideal subset that works reasonably well because they can't really leave such platform as they became hostages of their investment into it.

[0]: https://multiply.info/times-what-equals/4-times-what-equals-...

Maybe. The way I see it- whoever has the data wins in this new AI arms race. Google would have to make the biggest blunder in history to screw this up. It's not impossible but I wouldn't count them out. Their entire existence and practice of hoovering up all data has led to them this point.

> whoever has the data wins in this new AI arms race

This has been the running hypothesis, but it’s not panning out. Tesla, for example, doesn’t have the unambiguously best self-driving kit despite having unambiguously more data. Google has tons of data, but a lot of it is intelligently-tuned noise in the form of SEO spam.

>Tesla, for example, doesn’t have the unambiguously best self-driving kit despite having unambiguously more data.

Don’t they?

If we're being honest and not using backpedaling qualifiers like "available for purchase" or "designed without an ODD", then it's hard to argue how they could be. Multiple companies are operating driverless fleets in cities around the world. Tesla is not one of them.

We all “have the data”.

I build private models scraping the web. I’m iterating on my own virtual world generator using prompts, open source Vulkan renderer. Velocities, positions, color gradients; trivial to for loop your way to a massive DB with open libs.

Technologists are “screwing up” ogling Google/MS like passive consumer drones not seeing doing what Google and MS are doing is do-able at home with curl and open ML libs.

> and launch a half-baked product

It'll be fortunate if it's only half-baked. These digital assitants are already rough around the edges with how they sometimes confidently provide inaccurate information.

Add on Google's absolutely abysmal product development track record and the internal confusion over being forced into doing this and this isn't a stand alone application but it's being integrated into something, and it's a lot.

>These digital assitants are already rough around the edges with how they sometimes confidently provide inaccurate information.

its only going to get worse. because there will LLMs like chatgpt pumping out content for SEO content mills with no vetting for accuracy, then the next generation of LLMs are trained on them thus further corrupting the knowledge base they pull on to produce new content. its going to be a horrible feed back loop.

Yes! You're the first person I've heard say that and I'm sure you're right.

Then there will be some market for better-trained LLMs, which it will still be possible to make using curated data sets, but that will be expensive, and product differentiation will be messy. We're in for quite a ride. I can't predict the outcome. Maybe there is no stable outcome.

Indeed. We thought that the AI threat was killer robots. But it is actually the cancer of information which looks reputable but is incorrect in unpredictable ways.

"I do think there is unique break through here, because I feel that SEO has so thoroughly ruined search in many regards that this -could- be the right moment for this."

GPT-3 generally uses the same sources as search. Does GPT-3 using CommonCrawl have some way to exclude all the SEO garbage that is in the crawl data.

The difference with ChatGPT is there is no way to know that the output has been constructed from SEO garbage. Arguably that's even worse than SEO URLs that one can easily identify and avoid.

The comment about the press loving these projects is spot on. "Tech" employees love to criticise "mass media", except when mass media is promoting their (money-losing) projects.

> ... Google could overreact and launch a half-baked product

There's very low fraction of Google products that were sufficiently baked on release and just a few that weren't abandoned and killed after their half-baked release.

I wonder if Microsoft's Bing/Edge/ChatGPT integration will go any better than its attempt to marry NTFS and SQL Server (i.e. WinFS)? Spoiler: that didn't go as planned!

Was there any technological issue? Or was it just that developers didn't want to use it?

Microsoft never really revealed the 'real reasons' why it was canceled; so most people are left to speculate. My personal take is they decided it would become a tech-support nightmare to talk regular users through fixing a problem if the underlying database had an issue.

Here is an article from 2006: https://www.theguardian.com/technology/2006/jun/29/insideit....

They should be marketed as mascots, and given playful, childish personalities. You wouldn't blindly trust a child, why would you an AI in its infancy?

Clippy?

Microsoft really missed out by not having the bing AI assistant be clippy-based

I guarantee they focus-tested it and found that most people who remember Clippy don't have fond memories of him (and/or remember the memes/jokes about him more than they remember the actual Clippy), but, more importantly, most people don't remember him at all and wondered why this weird paper clip guy was talking. he hasn't been in an Office product since, what, 23 years ago or something? (not counting cameos)

And probably a fair number asked "what is this weird bent line thing and what does it have to do with searching the internet?"

"Okay, fine, we'll make him an anthropomorphic phone dial. Or maybe a floppy disc."

They trusted the "Red Queen". /s

Thank you. Now whenever I read output of a generative LLM, I will read it in my head with the confidently-wrong often-insane cadence of a toddler.

It will never happen because people will predictably abuse such bots in various gross ways, and the moment it leaks to the press, there will be a "think of the children" scandal.

> half-baked product

I'll make a 30,000 ft observation that LLMs, by definition, are half-baked bullshit generators. They can be useful, but they are full of warts.

as in vacuous memorizers, reciting things they've seen, sometimes in new combinations?

They definitely aren't vacuous memorizers. GP's description of them as "bullshit generators" is correct. They generate plausible-sounding text that (when it includes facts) is often counterfactual.

Much better said, i agree

I would love it if this nukes the SEO industry and the internet goes back to forums and community groups.

The best info is almost always locked up in these places.

LLMs will do the opposite. The internet will get flooded with machine-generated bullshit.

exactly. my only hope is that the flooding of cheap bulshit craters the market via oversupply. a combination of to much shit text competeing for ad space drops their revenue and people having diminishing trust for it reducing the ad clicks for it dropping the ad revenue even more

But LLM, while able to generate spam, if used used for search will reduce the need to click on a website, reducing the traffic to SEO websites… driving down profitability of doing it.

I feel like what I want is a LLM that can tell me not only a summary answer, but what search terms are most likely to return documentation for the answer. Provide clickable citations documenting the answer using those terms (it can iterate on the answer and terms if they aren't internally consistent). Then give summaries of the information at each citation and/or parameters to input if the response requires further operations/searches. It could provide some commentary of the quality of those citation/sources as well.

That makes the response more current, allows a further directed search based on the new parameters, and provides a traceable path to sources and citations. If it finds that there's not a Python library with that name then iterate.

Basically, an LLM should have a very good idea what good search terms are for the topic, and where to find the information, whereas I might not know the acronyms, jargon, or related fields.

Yes, this is getting pretty close to writing an 8th grade class report. That's about where these seem to be.

> lose these companies money

How can you say Siri loses Apple money? Does GarageBand lose them money? Photo Booth? Contacts?

By that logic Alexa can't lose Amazon any money because it's baked into their products. But in fact, it's a massive money suck. It lost $10B last year, and there's no way you can say Apple magically avoided losses, because Google had similar losses. There is a reason that none of the assistants have really improved as much as you'd expect over a decade+, it's not profitable, and not only that, but it's also -extremely- expensive. Microsoft basically just gave up on it; not really because of market share, but because it made no money, was expensive, and had limited value to customers.

https://arstechnica.com/gadgets/2022/11/amazon-alexa-is-a-co...

https://arstechnica.com/gadgets/2022/10/report-google-double...

https://www.theverge.com/22704233/siri-apple-digital-assista...

https://www.msn.com/en-us/news/technology/the-failure-of-ama...

How hard do you really think it is to build intents into something like Siri once you have the underlying technology framework?

Yes I have experience building intents on top of something like Siri.

It’s not the complexity of building intents that costs the money. Near real-time speech inference at scale doesn’t come for free. It’s only very recently that has started moving to the edge.

Speech inference has been done well enough locally for well over a decade. While an Alexa device probably couldn’t do it, any modern iPhone could.

> Speech inference has been done well enough locally for well over a decade

The first Alexa device was released in March 2014 - almost 9 years ago. Siri was a couple of years before that.

> While an Alexa device probably couldn’t do it, any modern iPhone could.

And that is why it is done in the cloud and has only recently started moving to the edge.

If the Echo devices were wildly profitable, earning $100+ billion per year, then I would say Alexa isn't losing money for Amazon.

Siri is not why people are buying iPhones, in fact well over 50% of users say they never or rarely use it. So, your point doesn't really hold up. Siri still loses money, just like Google Assistant and Alexa. The apps you listed to compare with OS bundles have not even close to the same overhead as a digital assistant and can be handled with relatively small programming teams.

Apple is selling HomePods and iPhones for a profit. But Amazon is/was selling Echos for a loss, in the hope that they would drive additional Prime music subscriptions and lower friction to Amazon purchases. That didn’t happen, and so the product line all-devices-sold-with-Alexa has lost Amazon money. The equivalent comparison, all devices with Siri, has made Apple a ton of money.

Without examining internal Amazon or Apple data on the marginal sales increase estimates for Siri/Alexa it's impossible to say, though.

Apple has earned a lot of money on devices with Siri included. Microsoft has made a lot of money on OSes with Minesweeper included.

It would be wrong to conclude anything about the profitability of Minesweeper based on that fact, just as it is wrong to conclude anything about the profitability of Siri based on the equivalent fact from Apple.

Apple's HomePods have been a resounding failure though? I don't know where you're getting your data, but it's wrong. They literally discontinued the original model because sales were SO bad.

Speaking of data, what are you basing that on? They seem to be selling well according to industry reports and companies tend not to continue products which lose money.

They've sold a billion dollars worth of HomePod minis. Even at Apple scale, that's not bad.

They've just launched a new HomePod that replaces the original. It's weird that there was such a big gap between the first one and the new one.

>Siri is not why people are buying iPhones, in fact well over 50% of users say they never or rarely use it

So? A decent portion of users do use it, and for some of them its a vital feature.

I would be shocked if removing siri would save Apple more money than the lost revenue of fewer iphone sales.

I love how everyone is missing the point that the assistants lose money. You can subsidize the loss, but it's still not profitable. People do not buy the iPhone because of Siri, especially given it is the worst of the big 3 assistants.

In management accounting, everything has a cost that has to be quantified, regardless of whether it's a standalone product or part of a HW/SW bundle. At the most basic levels, you have revenue centers (iPhone unit, iCloud unit), and cost centers (Apple Maps unit, customer support unit). All of these have operating costs.

The only way Siri does not lose Apple money is if there would be materially fewer iPhones sold if Siri was eliminated. In other words, if Siri is not a product differentiator, it's likely losing money (in the management accounting sense).

If you removed Siri, it would serious limit what you can do in CarPlay and you would also lose voice dictation. I think there would be materially fewer iPhones sold.

Although if apple didn't have Siri, then they would likely allow other voice assistants on the platform, and you'd have 'dictation by alexa' instead.

Agree about CarPlay. Siri is also pretty much indispensable for HomePods, though that's probably a minor product for Apple, and it plays an important role in AirPods.

Because they pay for servers and engineers for a product that they don’t charge for. It also doesn’t help sell phones when only 15% of users use it at least daily:

https://www.statista.com/statistics/696740/united-states-vir...

I find it incredibly hard to believe that a feature that 15% of users are using every day is not helping to sell phones.

Apple has something like 2 billion devices in use out there. 15% of that would be quite nice: https://www.fool.com/investing/2023/02/03/why-apple-stock-wa...

> Amazon Alexa, Google Assistant and Siri

These were always bullshit, other than maybe for turning your lights on, setting alarms, & checking the weather.

On the other hand, I immediately found ChatGPT / Copilot useful as a professional tool I could use to make my job measurably easier.

A minimum what could the be is voice interface to my phone apps and they can't even do that at all or in any consistent manner.

Agree. Only thing I want to add is that it's not just SEO ruining search, it's contradictory incentives. A user wants the quickest path to an answer but Google actually wants you to engage as long as possible so that you can look at ads.

If you'd now create a short/brief web page with the perfect answer to a particular question without any ads, Google would fully dismiss your page.

That's the tragedy of Google, never breaking away from that perversion.

> I think it could be game changing in many ways, but I can also see both of these falling apart.

Like that time the Microsoft Twitter chatbot got tricked into parroting white supremacist talking points within an afternoon?

Neeva has an AI for search. It works sufficiently well enough on some searches that it's all I'll read. Others I can tell it's missing things.

Measuring how useful something is by how much money it makes is like measuring a life by how intelligent it is.

Google Assistant and Siri were liked by the press, but not by me. In contrast, ChatGPT has helped me out more than a few times.

Same here. I treat ChatGPT as another Wikipedia or Stack Overflow. I know that the content is not fact checked by experts and I need to judge it accordingly. But just like Wikipedia can get you started on a topic, ChatGPT can do the same, plus you can ask follow up clarifying questions!

I wonder how AI SEO will look like. Bribe click workers to allow advertising to pass through in the ground truth data ?

One way I imagine this is that answers will have “helpful” postfixes after them. When the user interacts with the postfix prompts it serves pay-to-play dialogue. Advertisers can just bid on spots in a decision table. Simple example: a user asks “How do I best take care of roses in my location”? The ChatGPT/Bard answer is delivered then analyzed for keywords/themes that someone may want to advertise on. At the end of the answer (These results may be sponsored. Learn more) - Do you need help finding the nutrients mentioned in this response?

Maybe companies will add dozens of pages of text to their websites describing every little detail of their great products in every way possible in the hopes of getting picked up for the training data for the models. That text would of course also be written by some ai...

Thinking about how tech products are covered in the mainstream press, especially in the past when a lot of older journalists really didn’t have enough experience to ask the right questions, I’m betting this will be similar: write authoritative-sounding prose suggesting not just your company’s products but their preferred terminology and world-view, and spread it around as widely as possible. Get experts, real or self-labeled, to make ostensibly neutral buyer’s guides or reviews which conveniently exclude competitors (e.g. if my job was selling an Android phone, I might assert that nobody should even consider a closed-source OS for security reasons because that precludes the biggest decision while appearing to be objective).

Voice assistants probably don’t make them money but damn if I don’t use google assistant every day when I drive.

This does seem essentially like a Siri 2.0. I never found Siri useful but maybe this time it’ll be

siri sucks, but google assistant is already pretty good for the 90% of things I want it to do. a much improved google assistant would be awesome, and if I remember to use it within its limitations, I'm not worried about any of this "confidently wrong answers" stuff.

Reply on news.ycombinator.com